Abstract

Introduction: The aim of this study was to evaluate data mining techniques applying different filters in bone scan with ^99mTc-MDP and evaluate its performance to classify images into non-metastatic vs. metastatic.

Methods: We retrospectively studied 200 patients aged 45 to 85 years (mean 65 years) who underwent bone scans. The presence or absence of bone metastases was determined through a review of the reports from Nuclear Medicine physicians. The bone scans were processed in WEKA using automated machine learning software Auto-WEKA with RandomForest model and applying one or two filters to the images.

Results: Filter combinations showed the best performance to classify the scintigraphy into non-metastatic vs. metastatic, with a sensitivity of 100 vs. 98 %, specificity of 98% and areas under the ROC curve (AUC) of 100%.

Conclusion: The use of an automated machine learning system allows the classification of bone scans with high accuracy.

Key words: Auto-WEKA, data mining, artificial intelligence, bone scan, prostate cancer, image classification.